Two boxes at $749 reroute every AI call locally. The builder's $700 monthly cloud bill, cancelled. Jetson Orin Nano Super: NVIDIA spec, 8GB, 67 TOPS, $249 once. RTX 3090: 24GB VRAM, $500 used. Electricity: $6 on the Jetson, ~$10 on the 3090, the builder's figures. First year on the Jetson path: $321 total vs $8,400 in subscriptions. ROUTE - Ollama on the 3090, one env var reroutes Claude Code from the cloud to a local endpoint LOAD - qwen3-coder-30b fills the coding seat at 18GB SWAP - qwen3.8-27b takes chat at 18GB, devstral small 2 covers repos at 12GB SPLIT - Jetson handles decisions with a 4B model at 4.4GB and a 9B browser agent, zero API calls WIRE - phone app connects straight to the Jetson over local network, same UI as the cloud The 8GB board does not load a 35B model. A local model at 12–18GB is not the same product as a $200 cloud tier: narrower context, no managed tool layer. What it gives back is zero per-token cost and every prompt stays on the box. Full routing map in the terminal below ↓